Google DeepMind has introduced a new version of its generative AI for video creation — Veo 2. This version is the successor to Veo and can generate videos over two minutes long in up to 4K resolution. That’s four times the resolution and six times the duration of Sora from OpenAI. However, in the experimental VideoFX tool, where Veo 2 is currently available, videos are limited to 720p and eight seconds.
According to DeepMind’s VP of Product, Eli Collins, the company plans to gradually expand access to VideoFX and integrate Veo 2 into the Vertex AI developer platform. “In the coming months, we will continue to improve the model based on user feedback,” Collins noted.
Veo 2 can generate videos based on text prompts or a combination of text and images. The new version features improved understanding of physics and camera control, as well as enhanced image sharpness. Veo 2 is capable of realistically modeling motion, fluid dynamics, and light properties. However, the model still struggles with character consistency and details.
DeepMind continues to collaborate with artists and producers, including Donald Glover and The Weeknd, to improve video generation models. “We look forward to working with trusted testers and creators to gather feedback,” Collins stated.
To enhance safety, DeepMind uses SynthID watermarking technology to embed invisible markers into frames generated by Veo 2. At the same time, Google announced an update to its image generation model Imagen 3, which can now create more vibrant and detailed images in various styles.
